@Steve_Yegge: SkillBench is one of the most crazily important startups I know about, and it's been tough not to talk about them. Cong…
Summary
Steve Yegge praises SkillBench, a startup that scans coding agent session traces to build skills profiles and improve token efficiency. Matt Beane announces he's stepping in as full-time CEO.
View Cached Full Text
Cached at: 07/02/26, 06:19 AM
SkillBench is one of the most crazily important startups I know about, and it’s been tough not to talk about them. Congrats to @mattbeane on this huge move!
SkillBench is poised to solve a tremendous number of problems in the industry, not least of which could be token efficiency. SkillBench is really one of the most useful things I’ve ever seen come out of the AI era.
In short, and this is butchering it, they scan your coding agent session traces and build a skills profile from it. And you get a character sheet, of sorts. Life as RPG. Are you inefficient with tokens? They’ll help you see how, and even connect you to mentors.
They’re actually much more significant than I’ve hinted at here, but I figured I’d start with token efficiency since it’s arguably the industry’s biggest problem today. Everyone is hacking on token efficiency metrics with heuristics that try to tie outcomes to dev inputs. Only SkillBench treats the problem as a fine-grained skills discovery issue, and I think that means they’re the only ones who are doing it right.
I’ll be talking about them a lot more. I think they are huge piece of the puzzle for “How do organizations work in the future, in a way that helps humans flourish?”
Matt Beane (@mattbeane): For those who know me professionally, I’ll just steal the thunder from the end of this piece to make a clean announcement.
Today I go on academic leave, and start as full-time CEO of @skillbenchinc.
We are shipping what I talk about here, and more. Ignore our site. More soon.
Similar Articles
@gneubig: SkillsBench is a great benchmark, and I'm not just saying that because @OpenHandsDev beats everyone else on it Skills a…
SkillsBench 1.1 is a benchmark for evaluating AI agents' ability to use skills, now fully audited and error-free.
@xdotli: people came to our discord and ask how to write good proposal for making an eval when i started skillsbench, we have 0 …
SkillsBench founder shares the project's rapid growth from zero to 1600+ Discord members, 2 papers, and 150+ citations in under six months, along with extensive documentation.
@xdotli: A big pain point in using AI benchmarks is encountering errors after its first release. Today, we're releasing SkillsBe…
SkillsBench 1.1 is released as the first audited, error-free benchmark for AI agent skills, showing rapid capability improvement from ~36% to 67% resolution rate and demonstrating that skills can substitute for model scale.
@vincentsunnchen: New Benchtalks with @jyangballin: on ProgramBench (0% frontier models at launch) and the lineage/future of coding bench…
A podcast/interview episode discussing ProgramBench, a new coding benchmark where frontier models scored 0% at launch, covering its design philosophy, artifact-level evaluation, and the evolution of coding benchmarks from SWE-bench and InterCode.
@alex_prompter: Y Combinator CEO, Garry Tan, took the stage for 42 minutes at Startup School 2026 and explained how to build your own p…
Y Combinator CEO Garry Tan shares insights from his Startup School 2026 talk on building personal AGI, emphasizing that context and skill files matter more than model choice, and that capturing reusable skills can multiply output dramatically.